Classifying flows and buffer state for YouTube's HTTP adaptive streaming service in mobile networks

Accurate cross-layer information is very useful for optimizing and monitoring mobile networks for specific applications. However, the wide adoption of end-to-end encryption and the absence of common standards has made it very difficult to obtain such information by deep packet inspection (DPI) or cross-layer signaling. In this paper, we present a traffic profiling system as an alternative solution for HTTP adaptive streaming (HAS) traffic. By observing IP packet flows, our machine learning system detects the video flows as well as the play-back buffer state of an HAS client in real time. Studying 5 classification methods with an extensive dataset for YouTube’s mobile client, shows very high accuracy even with a strong variation of link quality. Since this high performance is achieved for a generic feature set at IP level, our approach requires no DPI, works equally with TCP and UDP, and does not interfere with end-to-end encryption. Traffic profiling is, thus, a powerful new tool for managing even encrypted HAS traffic in mobile networks.